net/ipv4/inet_fragment.c

   1 /*
   2  * inet fragments management
   3  *
   4  *              This program is free software; you can redistribute it and/or
   5  *              modify it under the terms of the GNU General Public License
   6  *              as published by the Free Software Foundation; either version
   7  *              2 of the License, or (at your option) any later version.
   8  *
   9  *              Authors:        Pavel Emelyanov <xemul@openvz.org>
  10  *                              Started as consolidation of ipv4/ip_fragment.c,
  11  *                              ipv6/reassembly. and ipv6 nf conntrack reassembly
  12  */
  13
  14 #include <linux/list.h>
  15 #include <linux/spinlock.h>
  16 #include <linux/module.h>
  17 #include <linux/timer.h>
  18 #include <linux/mm.h>
  19 #include <linux/random.h>
  20 #include <linux/skbuff.h>
  21 #include <linux/rtnetlink.h>
  22 #include <linux/slab.h>
  23
  24 #include <net/sock.h>
  25 #include <net/inet_frag.h>
  26 #include <net/inet_ecn.h>
  27
  28 #define INETFRAGS_EVICT_BUCKETS   128
  29 #define INETFRAGS_EVICT_MAX       512
  30
  31 /* don't rebuild inetfrag table with new secret more often than this */
  32 #define INETFRAGS_MIN_REBUILD_INTERVAL (5 * HZ)
  33
  34 /* Given the OR values of all fragments, apply RFC 3168 5.3 requirements
  35  * Value : 0xff if frame should be dropped.
  36  *         0 or INET_ECN_CE value, to be ORed in to final iph->tos field
  37  */
  38 const u8 ip_frag_ecn_table[16] = {
  39         /* at least one fragment had CE, and others ECT_0 or ECT_1 */
  40         [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0]                      = INET_ECN_CE,
  41         [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1]                      = INET_ECN_CE,
  42         [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1]   = INET_ECN_CE,
  43
  44         /* invalid combinations : drop frame */
  45         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE] = 0xff,
  46         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0] = 0xff,
  47         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_1] = 0xff,
  48         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff,
  49         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0] = 0xff,
  50         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1] = 0xff,
  51         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff,
  52 };
  53 EXPORT_SYMBOL(ip_frag_ecn_table);
  54
  55 static unsigned int
  56 inet_frag_hashfn(const struct inet_frags *f, const struct inet_frag_queue *q)
  57 {
  58         return f->hashfn(q) & (INETFRAGS_HASHSZ - 1);
  59 }
  60
  61 static bool inet_frag_may_rebuild(struct inet_frags *f)
  62 {
  63         return time_after(jiffies,
  64                f->last_rebuild_jiffies + INETFRAGS_MIN_REBUILD_INTERVAL);
  65 }
  66
  67 static void inet_frag_secret_rebuild(struct inet_frags *f)
  68 {
  69         int i;
  70
  71         write_seqlock_bh(&f->rnd_seqlock);
  72
  73         if (!inet_frag_may_rebuild(f))
  74                 goto out;
  75
  76         get_random_bytes(&f->rnd, sizeof(u32));
  77
  78         for (i = 0; i < INETFRAGS_HASHSZ; i++) {
  79                 struct inet_frag_bucket *hb;
  80                 struct inet_frag_queue *q;
  81                 struct hlist_node *n;
  82
  83                 hb = &f->hash[i];
  84                 spin_lock(&hb->chain_lock);
  85
  86                 hlist_for_each_entry_safe(q, n, &hb->chain, list) {
  87                         unsigned int hval = inet_frag_hashfn(f, q);
  88
  89                         if (hval != i) {
  90                                 struct inet_frag_bucket *hb_dest;
  91
  92                                 hlist_del(&q->list);
  93
  94                                 /* Relink to new hash chain. */
  95                                 hb_dest = &f->hash[hval];
  96
  97                                 /* This is the only place where we take
  98                                  * another chain_lock while already holding
  99                                  * one.  As this will not run concurrently,
 100                                  * we cannot deadlock on hb_dest lock below, if its
 101                                  * already locked it will be released soon since
 102                                  * other caller cannot be waiting for hb lock
 103                                  * that we've taken above.
 104                                  */
 105                                 spin_lock_nested(&hb_dest->chain_lock,
 106                                                  SINGLE_DEPTH_NESTING);
 107                                 hlist_add_head(&q->list, &hb_dest->chain);
 108                                 spin_unlock(&hb_dest->chain_lock);
 109                         }
 110                 }
 111                 spin_unlock(&hb->chain_lock);
 112         }
 113
 114         f->rebuild = false;
 115         f->last_rebuild_jiffies = jiffies;
 116 out:
 117         write_sequnlock_bh(&f->rnd_seqlock);
 118 }
 119
 120 static bool inet_fragq_should_evict(const struct inet_frag_queue *q)
 121 {
 122         if (!hlist_unhashed(&q->list_evictor))
 123                 return false;
 124
 125         return q->net->low_thresh == 0 ||
 126                frag_mem_limit(q->net) >= q->net->low_thresh;
 127 }
 128
 129 static unsigned int
 130 inet_evict_bucket(struct inet_frags *f, struct inet_frag_bucket *hb)
 131 {
 132         struct inet_frag_queue *fq;
 133         struct hlist_node *n;
 134         unsigned int evicted = 0;
 135         HLIST_HEAD(expired);
 136
 137         spin_lock(&hb->chain_lock);
 138
 139         hlist_for_each_entry_safe(fq, n, &hb->chain, list) {
 140                 if (!inet_fragq_should_evict(fq))
 141                         continue;
 142
 143                 if (!del_timer(&fq->timer))
 144                         continue;
 145
 146                 hlist_add_head(&fq->list_evictor, &expired);
 147                 ++evicted;
 148         }
 149
 150         spin_unlock(&hb->chain_lock);
 151
 152         hlist_for_each_entry_safe(fq, n, &expired, list_evictor)
 153                 f->frag_expire((unsigned long) fq);
 154
 155         return evicted;
 156 }
 157
 158 static void inet_frag_worker(struct work_struct *work)
 159 {
 160         unsigned int budget = INETFRAGS_EVICT_BUCKETS;
 161         unsigned int i, evicted = 0;
 162         struct inet_frags *f;
 163
 164         f = container_of(work, struct inet_frags, frags_work);
 165
 166         BUILD_BUG_ON(INETFRAGS_EVICT_BUCKETS >= INETFRAGS_HASHSZ);
 167
 168         local_bh_disable();
 169
 170         for (i = ACCESS_ONCE(f->next_bucket); budget; --budget) {
 171                 evicted += inet_evict_bucket(f, &f->hash[i]);
 172                 i = (i + 1) & (INETFRAGS_HASHSZ - 1);
 173                 if (evicted > INETFRAGS_EVICT_MAX)
 174                         break;
 175         }
 176
 177         f->next_bucket = i;
 178
 179         local_bh_enable();
 180
 181         if (f->rebuild && inet_frag_may_rebuild(f))
 182                 inet_frag_secret_rebuild(f);
 183 }
 184
 185 static void inet_frag_schedule_worker(struct inet_frags *f)
 186 {
 187         if (unlikely(!work_pending(&f->frags_work)))
 188                 schedule_work(&f->frags_work);
 189 }
 190
 191 int inet_frags_init(struct inet_frags *f)
 192 {
 193         int i;
 194
 195         INIT_WORK(&f->frags_work, inet_frag_worker);
 196
 197         for (i = 0; i < INETFRAGS_HASHSZ; i++) {
 198                 struct inet_frag_bucket *hb = &f->hash[i];
 199
 200                 spin_lock_init(&hb->chain_lock);
 201                 INIT_HLIST_HEAD(&hb->chain);
 202         }
 203
 204         seqlock_init(&f->rnd_seqlock);
 205         f->last_rebuild_jiffies = 0;
 206         f->frags_cachep = kmem_cache_create(f->frags_cache_name, f->qsize, 0, 0,
 207                                             NULL);
 208         if (!f->frags_cachep)
 209                 return -ENOMEM;
 210
 211         return 0;
 212 }
 213 EXPORT_SYMBOL(inet_frags_init);
 214
 215 void inet_frags_fini(struct inet_frags *f)
 216 {
 217         cancel_work_sync(&f->frags_work);
 218         kmem_cache_destroy(f->frags_cachep);
 219 }
 220 EXPORT_SYMBOL(inet_frags_fini);
 221
 222 void inet_frags_exit_net(struct netns_frags *nf, struct inet_frags *f)
 223 {
 224         unsigned int seq;
 225         int i;
 226
 227         nf->low_thresh = 0;
 228
 229 evict_again:
 230         local_bh_disable();
 231         seq = read_seqbegin(&f->rnd_seqlock);
 232
 233         for (i = 0; i < INETFRAGS_HASHSZ ; i++)
 234                 inet_evict_bucket(f, &f->hash[i]);
 235
 236         local_bh_enable();
 237         cond_resched();
 238
 239         if (read_seqretry(&f->rnd_seqlock, seq) ||
 240             sum_frag_mem_limit(nf))
 241                 goto evict_again;
 242 }
 243 EXPORT_SYMBOL(inet_frags_exit_net);
 244
 245 static struct inet_frag_bucket *
 246 get_frag_bucket_locked(struct inet_frag_queue *fq, struct inet_frags *f)
 247 __acquires(hb->chain_lock)
 248 {
 249         struct inet_frag_bucket *hb;
 250         unsigned int seq, hash;
 251
 252  restart:
 253         seq = read_seqbegin(&f->rnd_seqlock);
 254
 255         hash = inet_frag_hashfn(f, fq);
 256         hb = &f->hash[hash];
 257
 258         spin_lock(&hb->chain_lock);
 259         if (read_seqretry(&f->rnd_seqlock, seq)) {
 260                 spin_unlock(&hb->chain_lock);
 261                 goto restart;
 262         }
 263
 264         return hb;
 265 }
 266
 267 static inline void fq_unlink(struct inet_frag_queue *fq, struct inet_frags *f)
 268 {
 269         struct inet_frag_bucket *hb;
 270
 271         hb = get_frag_bucket_locked(fq, f);
 272         hlist_del(&fq->list);
 273         fq->flags |= INET_FRAG_COMPLETE;
 274         spin_unlock(&hb->chain_lock);
 275 }
 276
 277 void inet_frag_kill(struct inet_frag_queue *fq, struct inet_frags *f)
 278 {
 279         if (del_timer(&fq->timer))
 280                 atomic_dec(&fq->refcnt);
 281
 282         if (!(fq->flags & INET_FRAG_COMPLETE)) {
 283                 fq_unlink(fq, f);
 284                 atomic_dec(&fq->refcnt);
 285         }
 286 }
 287 EXPORT_SYMBOL(inet_frag_kill);
 288
 289 static inline void frag_kfree_skb(struct netns_frags *nf, struct inet_frags *f,
 290                                   struct sk_buff *skb)
 291 {
 292         if (f->skb_free)
 293                 f->skb_free(skb);
 294         kfree_skb(skb);
 295 }
 296
 297 void inet_frag_destroy(struct inet_frag_queue *q, struct inet_frags *f)
 298 {
 299         struct sk_buff *fp;
 300         struct netns_frags *nf;
 301         unsigned int sum, sum_truesize = 0;
 302
 303         WARN_ON(!(q->flags & INET_FRAG_COMPLETE));
 304         WARN_ON(del_timer(&q->timer) != 0);
 305
 306         /* Release all fragment data. */
 307         fp = q->fragments;
 308         nf = q->net;
 309         while (fp) {
 310                 struct sk_buff *xp = fp->next;
 311
 312                 sum_truesize += fp->truesize;
 313                 frag_kfree_skb(nf, f, fp);
 314                 fp = xp;
 315         }
 316         sum = sum_truesize + f->qsize;
 317
 318         if (f->destructor)
 319                 f->destructor(q);
 320         kmem_cache_free(f->frags_cachep, q);
 321
 322         sub_frag_mem_limit(nf, sum);
 323 }
 324 EXPORT_SYMBOL(inet_frag_destroy);
 325
 326 static struct inet_frag_queue *inet_frag_intern(struct netns_frags *nf,
 327                                                 struct inet_frag_queue *qp_in,
 328                                                 struct inet_frags *f,
 329                                                 void *arg)
 330 {
 331         struct inet_frag_bucket *hb = get_frag_bucket_locked(qp_in, f);
 332         struct inet_frag_queue *qp;
 333
 334 #ifdef CONFIG_SMP
 335         /* With SMP race we have to recheck hash table, because
 336          * such entry could have been created on other cpu before
 337          * we acquired hash bucket lock.
 338          */
 339         hlist_for_each_entry(qp, &hb->chain, list) {
 340                 if (qp->net == nf && f->match(qp, arg)) {
 341                         atomic_inc(&qp->refcnt);
 342                         spin_unlock(&hb->chain_lock);
 343                         qp_in->flags |= INET_FRAG_COMPLETE;
 344                         inet_frag_put(qp_in, f);
 345                         return qp;
 346                 }
 347         }
 348 #endif
 349         qp = qp_in;
 350         if (!mod_timer(&qp->timer, jiffies + nf->timeout))
 351                 atomic_inc(&qp->refcnt);
 352
 353         atomic_inc(&qp->refcnt);
 354         hlist_add_head(&qp->list, &hb->chain);
 355
 356         spin_unlock(&hb->chain_lock);
 357
 358         return qp;
 359 }
 360
 361 static struct inet_frag_queue *inet_frag_alloc(struct netns_frags *nf,
 362                                                struct inet_frags *f,
 363                                                void *arg)
 364 {
 365         struct inet_frag_queue *q;
 366
 367         if (frag_mem_limit(nf) > nf->high_thresh) {
 368                 inet_frag_schedule_worker(f);
 369                 return NULL;
 370         }
 371
 372         q = kmem_cache_zalloc(f->frags_cachep, GFP_ATOMIC);
 373         if (!q)
 374                 return NULL;
 375
 376         q->net = nf;
 377         f->constructor(q, arg);
 378         add_frag_mem_limit(nf, f->qsize);
 379
 380         setup_timer(&q->timer, f->frag_expire, (unsigned long)q);
 381         spin_lock_init(&q->lock);
 382         atomic_set(&q->refcnt, 1);
 383
 384         return q;
 385 }
 386
 387 static struct inet_frag_queue *inet_frag_create(struct netns_frags *nf,
 388                                                 struct inet_frags *f,
 389                                                 void *arg)
 390 {
 391         struct inet_frag_queue *q;
 392
 393         q = inet_frag_alloc(nf, f, arg);
 394         if (!q)
 395                 return NULL;
 396
 397         return inet_frag_intern(nf, q, f, arg);
 398 }
 399
 400 struct inet_frag_queue *inet_frag_find(struct netns_frags *nf,
 401                                        struct inet_frags *f, void *key,
 402                                        unsigned int hash)
 403 {
 404         struct inet_frag_bucket *hb;
 405         struct inet_frag_queue *q;
 406         int depth = 0;
 407
 408         if (frag_mem_limit(nf) > nf->low_thresh)
 409                 inet_frag_schedule_worker(f);
 410
 411         hash &= (INETFRAGS_HASHSZ - 1);
 412         hb = &f->hash[hash];
 413
 414         spin_lock(&hb->chain_lock);
 415         hlist_for_each_entry(q, &hb->chain, list) {
 416                 if (q->net == nf && f->match(q, key)) {
 417                         atomic_inc(&q->refcnt);
 418                         spin_unlock(&hb->chain_lock);
 419                         return q;
 420                 }
 421                 depth++;
 422         }
 423         spin_unlock(&hb->chain_lock);
 424
 425         if (depth <= INETFRAGS_MAXDEPTH)
 426                 return inet_frag_create(nf, f, key);
 427
 428         if (inet_frag_may_rebuild(f)) {
 429                 if (!f->rebuild)
 430                         f->rebuild = true;
 431                 inet_frag_schedule_worker(f);
 432         }
 433
 434         return ERR_PTR(-ENOBUFS);
 435 }
 436 EXPORT_SYMBOL(inet_frag_find);
 437
 438 void inet_frag_maybe_warn_overflow(struct inet_frag_queue *q,
 439                                    const char *prefix)
 440 {
 441         static const char msg[] = "inet_frag_find: Fragment hash bucket"
 442                 " list length grew over limit " __stringify(INETFRAGS_MAXDEPTH)
 443                 ". Dropping fragment.\n";
 444
 445         if (PTR_ERR(q) == -ENOBUFS)
 446                 net_dbg_ratelimited("%s%s", prefix, msg);
 447 }
 448 EXPORT_SYMBOL(inet_frag_maybe_warn_overflow);